Reduce WebGPU burst submissions (p99 5.3ms, 27 bursts)

[ysdede]

Trace: Trace-20260213T014231.json (summary in trace_analysis_summary.json)

Evidence

• webgpu: 2258 events over 22.5s (~100.3/s)
• p99 = 5358us, max = 10026us
• burst_count = 27, burst_size avg 8.9 (min 6, max 19)

Impact

• Bursty GPU submissions can stall rendering and increase tail latency during transcription.

Actions

• Add WebGPU markers around model inference to identify submit batches.
• Increase batch size / reduce submission frequency if possible (group work per inference tick).
• Verify if any UI rendering uses WebGPU concurrently and separate queues if supported.

Acceptance

• p99 WebGPU event duration < 3ms.
• Burst count reduced by 50% under same workload.

[ysdede]

From Feb 2026 trace. WebGPU submission patterns may have changed with the v4 utterance pipeline. Needs fresh profiling.